xref: /freebsd/sys/netinet/toecore.c (revision 206b73d0) 
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2012 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 
36 #include <sys/param.h>
37 #include <sys/eventhandler.h>
38 #include <sys/kernel.h>
39 #include <sys/systm.h>
40 #include <sys/malloc.h>
41 #include <sys/mbuf.h>
42 #include <sys/module.h>
43 #include <sys/types.h>
44 #include <sys/sockopt.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 
48 #include <net/ethernet.h>
49 #include <net/if.h>
50 #include <net/if_var.h>
51 #include <net/if_types.h>
52 #include <net/if_vlan_var.h>
53 #include <net/if_llatbl.h>
54 #include <net/route.h>
55 
56 #include <netinet/if_ether.h>
57 #include <netinet/in.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/in_var.h>
60 #include <netinet6/in6_var.h>
61 #include <netinet6/in6_pcb.h>
62 #include <netinet6/nd6.h>
63 #define TCPSTATES
64 #include <netinet/tcp.h>
65 #include <netinet/tcp_fsm.h>
66 #include <netinet/tcp_timer.h>
67 #include <netinet/tcp_var.h>
68 #include <netinet/tcp_syncache.h>
69 #include <netinet/tcp_offload.h>
70 #include <netinet/toecore.h>
71 
72 static struct mtx toedev_lock;
73 static TAILQ_HEAD(, toedev) toedev_list;
74 static eventhandler_tag listen_start_eh;
75 static eventhandler_tag listen_stop_eh;
76 static eventhandler_tag lle_event_eh;
77 
78 static int
79 toedev_connect(struct toedev *tod __unused, struct socket *so __unused,
80     struct rtentry *rt __unused, struct sockaddr *nam __unused)
81 {
82 
83 	return (ENOTSUP);
84 }
85 
86 static int
87 toedev_listen_start(struct toedev *tod __unused, struct tcpcb *tp __unused)
88 {
89 
90 	return (ENOTSUP);
91 }
92 
93 static int
94 toedev_listen_stop(struct toedev *tod __unused, struct tcpcb *tp __unused)
95 {
96 
97 	return (ENOTSUP);
98 }
99 
100 static void
101 toedev_input(struct toedev *tod __unused, struct tcpcb *tp __unused,
102     struct mbuf *m)
103 {
104 
105 	m_freem(m);
106 	return;
107 }
108 
109 static void
110 toedev_rcvd(struct toedev *tod __unused, struct tcpcb *tp __unused)
111 {
112 
113 	return;
114 }
115 
116 static int
117 toedev_output(struct toedev *tod __unused, struct tcpcb *tp __unused)
118 {
119 
120 	return (ENOTSUP);
121 }
122 
123 static void
124 toedev_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp __unused)
125 {
126 
127 	return;
128 }
129 
130 static void
131 toedev_l2_update(struct toedev *tod __unused, struct ifnet *ifp __unused,
132     struct sockaddr *sa __unused, uint8_t *lladdr __unused,
133     uint16_t vtag __unused)
134 {
135 
136 	return;
137 }
138 
139 static void
140 toedev_route_redirect(struct toedev *tod __unused, struct ifnet *ifp __unused,
141     struct rtentry *rt0 __unused, struct rtentry *rt1 __unused)
142 {
143 
144 	return;
145 }
146 
147 static void
148 toedev_syncache_added(struct toedev *tod __unused, void *ctx __unused)
149 {
150 
151 	return;
152 }
153 
154 static void
155 toedev_syncache_removed(struct toedev *tod __unused, void *ctx __unused)
156 {
157 
158 	return;
159 }
160 
161 static int
162 toedev_syncache_respond(struct toedev *tod __unused, void *ctx __unused,
163     struct mbuf *m)
164 {
165 
166 	m_freem(m);
167 	return (0);
168 }
169 
170 static void
171 toedev_offload_socket(struct toedev *tod __unused, void *ctx __unused,
172     struct socket *so __unused)
173 {
174 
175 	return;
176 }
177 
178 static void
179 toedev_ctloutput(struct toedev *tod __unused, struct tcpcb *tp __unused,
180     int sopt_dir __unused, int sopt_name __unused)
181 {
182 
183 	return;
184 }
185 
186 static void
187 toedev_tcp_info(struct toedev *tod __unused, struct tcpcb *tp __unused,
188     struct tcp_info *ti __unused)
189 {
190 
191 	return;
192 }
193 
194 /*
195  * Inform one or more TOE devices about a listening socket.
196  */
197 static void
198 toe_listen_start(struct inpcb *inp, void *arg)
199 {
200 	struct toedev *t, *tod;
201 	struct tcpcb *tp;
202 
203 	INP_WLOCK_ASSERT(inp);
204 	KASSERT(inp->inp_pcbinfo == &V_tcbinfo,
205 	    ("%s: inp is not a TCP inp", __func__));
206 
207 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
208 		return;
209 
210 	tp = intotcpcb(inp);
211 	if (tp->t_state != TCPS_LISTEN)
212 		return;
213 
214 	t = arg;
215 	mtx_lock(&toedev_lock);
216 	TAILQ_FOREACH(tod, &toedev_list, link) {
217 		if (t == NULL || t == tod)
218 			tod->tod_listen_start(tod, tp);
219 	}
220 	mtx_unlock(&toedev_lock);
221 }
222 
223 static void
224 toe_listen_start_event(void *arg __unused, struct tcpcb *tp)
225 {
226 	struct inpcb *inp = tp->t_inpcb;
227 
228 	INP_WLOCK_ASSERT(inp);
229 	KASSERT(tp->t_state == TCPS_LISTEN,
230 	    ("%s: t_state %s", __func__, tcpstates[tp->t_state]));
231 
232 	toe_listen_start(inp, NULL);
233 }
234 
235 static void
236 toe_listen_stop_event(void *arg __unused, struct tcpcb *tp)
237 {
238 	struct toedev *tod;
239 #ifdef INVARIANTS
240 	struct inpcb *inp = tp->t_inpcb;
241 #endif
242 
243 	INP_WLOCK_ASSERT(inp);
244 	KASSERT(tp->t_state == TCPS_LISTEN,
245 	    ("%s: t_state %s", __func__, tcpstates[tp->t_state]));
246 
247 	mtx_lock(&toedev_lock);
248 	TAILQ_FOREACH(tod, &toedev_list, link)
249 	    tod->tod_listen_stop(tod, tp);
250 	mtx_unlock(&toedev_lock);
251 }
252 
253 /*
254  * Fill up a freshly allocated toedev struct with reasonable defaults.
255  */
256 void
257 init_toedev(struct toedev *tod)
258 {
259 
260 	tod->tod_softc = NULL;
261 
262 	/*
263 	 * Provide no-op defaults so that the kernel can call any toedev
264 	 * function without having to check whether the TOE driver supplied one
265 	 * or not.
266 	 */
267 	tod->tod_connect = toedev_connect;
268 	tod->tod_listen_start = toedev_listen_start;
269 	tod->tod_listen_stop = toedev_listen_stop;
270 	tod->tod_input = toedev_input;
271 	tod->tod_rcvd = toedev_rcvd;
272 	tod->tod_output = toedev_output;
273 	tod->tod_send_rst = toedev_output;
274 	tod->tod_send_fin = toedev_output;
275 	tod->tod_pcb_detach = toedev_pcb_detach;
276 	tod->tod_l2_update = toedev_l2_update;
277 	tod->tod_route_redirect = toedev_route_redirect;
278 	tod->tod_syncache_added = toedev_syncache_added;
279 	tod->tod_syncache_removed = toedev_syncache_removed;
280 	tod->tod_syncache_respond = toedev_syncache_respond;
281 	tod->tod_offload_socket = toedev_offload_socket;
282 	tod->tod_ctloutput = toedev_ctloutput;
283 	tod->tod_tcp_info = toedev_tcp_info;
284 }
285 
286 /*
287  * Register an active TOE device with the system.  This allows it to receive
288  * notifications from the kernel.
289  */
290 int
291 register_toedev(struct toedev *tod)
292 {
293 	struct toedev *t;
294 
295 	mtx_lock(&toedev_lock);
296 	TAILQ_FOREACH(t, &toedev_list, link) {
297 		if (t == tod) {
298 			mtx_unlock(&toedev_lock);
299 			return (EEXIST);
300 		}
301 	}
302 
303 	TAILQ_INSERT_TAIL(&toedev_list, tod, link);
304 	registered_toedevs++;
305 	mtx_unlock(&toedev_lock);
306 
307 	inp_apply_all(toe_listen_start, tod);
308 
309 	return (0);
310 }
311 
312 /*
313  * Remove the TOE device from the global list of active TOE devices.  It is the
314  * caller's responsibility to ensure that the TOE device is quiesced prior to
315  * this call.
316  */
317 int
318 unregister_toedev(struct toedev *tod)
319 {
320 	struct toedev *t, *t2;
321 	int rc = ENODEV;
322 
323 	mtx_lock(&toedev_lock);
324 	TAILQ_FOREACH_SAFE(t, &toedev_list, link, t2) {
325 		if (t == tod) {
326 			TAILQ_REMOVE(&toedev_list, tod, link);
327 			registered_toedevs--;
328 			rc = 0;
329 			break;
330 		}
331 	}
332 	KASSERT(registered_toedevs >= 0,
333 	    ("%s: registered_toedevs (%d) < 0", __func__, registered_toedevs));
334 	mtx_unlock(&toedev_lock);
335 	return (rc);
336 }
337 
338 void
339 toe_syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
340     struct inpcb *inp, void *tod, void *todctx)
341 {
342 	struct socket *lso = inp->inp_socket;
343 
344 	INP_WLOCK_ASSERT(inp);
345 
346 	syncache_add(inc, to, th, inp, &lso, NULL, tod, todctx);
347 }
348 
349 int
350 toe_syncache_expand(struct in_conninfo *inc, struct tcpopt *to,
351     struct tcphdr *th, struct socket **lsop)
352 {
353 
354 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
355 
356 	return (syncache_expand(inc, to, th, lsop, NULL));
357 }
358 
359 /*
360  * General purpose check to see if a 4-tuple is in use by the kernel.  If a TCP
361  * header (presumably for an incoming SYN) is also provided, an existing 4-tuple
362  * in TIME_WAIT may be assassinated freeing it up for re-use.
363  *
364  * Note that the TCP header must have been run through tcp_fields_to_host() or
365  * equivalent.
366  */
367 int
368 toe_4tuple_check(struct in_conninfo *inc, struct tcphdr *th, struct ifnet *ifp)
369 {
370 	struct inpcb *inp;
371 
372 	if (inc->inc_flags & INC_ISIPV6) {
373 		inp = in6_pcblookup(&V_tcbinfo, &inc->inc6_faddr,
374 		    inc->inc_fport, &inc->inc6_laddr, inc->inc_lport,
375 		    INPLOOKUP_WLOCKPCB, ifp);
376 	} else {
377 		inp = in_pcblookup(&V_tcbinfo, inc->inc_faddr, inc->inc_fport,
378 		    inc->inc_laddr, inc->inc_lport, INPLOOKUP_WLOCKPCB, ifp);
379 	}
380 	if (inp != NULL) {
381 		INP_WLOCK_ASSERT(inp);
382 
383 		if ((inp->inp_flags & INP_TIMEWAIT) && th != NULL) {
384 
385 			INP_INFO_RLOCK_ASSERT(&V_tcbinfo); /* for twcheck */
386 			if (!tcp_twcheck(inp, NULL, th, NULL, 0))
387 				return (EADDRINUSE);
388 		} else {
389 			INP_WUNLOCK(inp);
390 			return (EADDRINUSE);
391 		}
392 	}
393 
394 	return (0);
395 }
396 
397 static void
398 toe_lle_event(void *arg __unused, struct llentry *lle, int evt)
399 {
400 	struct toedev *tod;
401 	struct ifnet *ifp;
402 	struct sockaddr *sa;
403 	uint8_t *lladdr;
404 	uint16_t vid, pcp;
405 	int family;
406 	struct sockaddr_in6 sin6;
407 
408 	LLE_WLOCK_ASSERT(lle);
409 
410 	ifp = lltable_get_ifp(lle->lle_tbl);
411 	family = lltable_get_af(lle->lle_tbl);
412 
413 	if (family != AF_INET && family != AF_INET6)
414 		return;
415 	/*
416 	 * Not interested if the interface's TOE capability is not enabled.
417 	 */
418 	if ((family == AF_INET && !(ifp->if_capenable & IFCAP_TOE4)) ||
419 	    (family == AF_INET6 && !(ifp->if_capenable & IFCAP_TOE6)))
420 		return;
421 
422 	tod = TOEDEV(ifp);
423 	if (tod == NULL)
424 		return;
425 
426 	sa = (struct sockaddr *)&sin6;
427 	lltable_fill_sa_entry(lle, sa);
428 
429 	vid = 0xfff;
430 	pcp = 0;
431 	if (evt != LLENTRY_RESOLVED) {
432 
433 		/*
434 		 * LLENTRY_TIMEDOUT, LLENTRY_DELETED, LLENTRY_EXPIRED all mean
435 		 * this entry is going to be deleted.
436 		 */
437 
438 		lladdr = NULL;
439 	} else {
440 
441 		KASSERT(lle->la_flags & LLE_VALID,
442 		    ("%s: %p resolved but not valid?", __func__, lle));
443 
444 		lladdr = (uint8_t *)lle->ll_addr;
445 		VLAN_TAG(ifp, &vid);
446 		VLAN_PCP(ifp, &pcp);
447 	}
448 
449 	tod->tod_l2_update(tod, ifp, sa, lladdr, EVL_MAKETAG(vid, pcp, 0));
450 }
451 
452 /*
453  * Returns 0 or EWOULDBLOCK on success (any other value is an error).  0 means
454  * lladdr and vtag are valid on return, EWOULDBLOCK means the TOE driver's
455  * tod_l2_update will be called later, when the entry is resolved or times out.
456  */
457 int
458 toe_l2_resolve(struct toedev *tod, struct ifnet *ifp, struct sockaddr *sa,
459     uint8_t *lladdr, uint16_t *vtag)
460 {
461 	int rc;
462 	uint16_t vid, pcp;
463 
464 	switch (sa->sa_family) {
465 #ifdef INET
466 	case AF_INET:
467 		rc = arpresolve(ifp, 0, NULL, sa, lladdr, NULL, NULL);
468 		break;
469 #endif
470 #ifdef INET6
471 	case AF_INET6:
472 		rc = nd6_resolve(ifp, 0, NULL, sa, lladdr, NULL, NULL);
473 		break;
474 #endif
475 	default:
476 		return (EPROTONOSUPPORT);
477 	}
478 
479 	if (rc == 0) {
480 		vid = 0xfff;
481 		pcp = 0;
482 		if (ifp->if_type == IFT_L2VLAN) {
483 			VLAN_TAG(ifp, &vid);
484 			VLAN_PCP(ifp, &pcp);
485 		} else if (ifp->if_pcp != IFNET_PCP_NONE) {
486 			vid = 0;
487 			pcp = ifp->if_pcp;
488 		}
489 		*vtag = EVL_MAKETAG(vid, pcp, 0);
490 	}
491 
492 	return (rc);
493 }
494 
495 void
496 toe_connect_failed(struct toedev *tod, struct inpcb *inp, int err)
497 {
498 
499 	INP_WLOCK_ASSERT(inp);
500 
501 	if (!(inp->inp_flags & INP_DROPPED)) {
502 		struct tcpcb *tp = intotcpcb(inp);
503 
504 		KASSERT(tp->t_flags & TF_TOE,
505 		    ("%s: tp %p not offloaded.", __func__, tp));
506 
507 		if (err == EAGAIN) {
508 
509 			/*
510 			 * Temporary failure during offload, take this PCB back.
511 			 * Detach from the TOE driver and do the rest of what
512 			 * TCP's pru_connect would have done if the connection
513 			 * wasn't offloaded.
514 			 */
515 
516 			tod->tod_pcb_detach(tod, tp);
517 			KASSERT(!(tp->t_flags & TF_TOE),
518 			    ("%s: tp %p still offloaded.", __func__, tp));
519 			tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
520 			(void) tp->t_fb->tfb_tcp_output(tp);
521 		} else {
522 
523 			INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
524 			tp = tcp_drop(tp, err);
525 			if (tp == NULL)
526 				INP_WLOCK(inp);	/* re-acquire */
527 		}
528 	}
529 	INP_WLOCK_ASSERT(inp);
530 }
531 
532 static int
533 toecore_load(void)
534 {
535 
536 	mtx_init(&toedev_lock, "toedev lock", NULL, MTX_DEF);
537 	TAILQ_INIT(&toedev_list);
538 
539 	listen_start_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_start,
540 	    toe_listen_start_event, NULL, EVENTHANDLER_PRI_ANY);
541 	listen_stop_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_stop,
542 	    toe_listen_stop_event, NULL, EVENTHANDLER_PRI_ANY);
543 	lle_event_eh = EVENTHANDLER_REGISTER(lle_event, toe_lle_event, NULL,
544 	    EVENTHANDLER_PRI_ANY);
545 
546 	return (0);
547 }
548 
549 static int
550 toecore_unload(void)
551 {
552 
553 	mtx_lock(&toedev_lock);
554 	if (!TAILQ_EMPTY(&toedev_list)) {
555 		mtx_unlock(&toedev_lock);
556 		return (EBUSY);
557 	}
558 
559 	EVENTHANDLER_DEREGISTER(tcp_offload_listen_start, listen_start_eh);
560 	EVENTHANDLER_DEREGISTER(tcp_offload_listen_stop, listen_stop_eh);
561 	EVENTHANDLER_DEREGISTER(lle_event, lle_event_eh);
562 
563 	mtx_unlock(&toedev_lock);
564 	mtx_destroy(&toedev_lock);
565 
566 	return (0);
567 }
568 
569 static int
570 toecore_mod_handler(module_t mod, int cmd, void *arg)
571 {
572 
573 	if (cmd == MOD_LOAD)
574 		return (toecore_load());
575 
576 	if (cmd == MOD_UNLOAD)
577 		return (toecore_unload());
578 
579 	return (EOPNOTSUPP);
580 }
581 
582 static moduledata_t mod_data= {
583 	"toecore",
584 	toecore_mod_handler,
585 	0
586 };
587 
588 MODULE_VERSION(toecore, 1);
589 DECLARE_MODULE(toecore, mod_data, SI_SUB_EXEC, SI_ORDER_ANY);
590